Telecom data shows characters that containing missing data rate

large difference in feature dimensions

high correlation in features

which current feature mining technology can not fully applied to. Therefore

a novel feature mining method was designed to deal with such problems

which was consist of four indicators

namely

the missing data rate

standard deviation

correlation and importance. Firstly

based on these indicators

different evaluation functions were fitted. Then the comprehensive score of each feature was calculated with weight value for feature mining. On the other hand

based on actual telecom data

comparing with Random Forest

which was one of the most commonly used feature mining algorithm

the conclusion was drawn that this novel method can obtain a more scienti?c and reasonable result for measuring from more dimensions

what’s more

it saves 25% cost of time and provides strong practicability.